EvoComp: Learning Visual Token Compression for Multimodal Large Language Models via Semantic-Guided Evolutionary Labeling

The key wrinkle here is the 'evolutionary' part: rather than using fixed human-labeled training data to teach the compressor which tokens matter, EvoComp iteratively refines its own labels during training, letting the model discover which visual regions are semantically load-bearing for a given text query. That self-supervised labeling loop is what separates this from earlier static compression approaches.

Token compression is getting crowded fast. Just two days before this paper dropped, we covered K-Token Merging (arXiv, April 16), which tackles the same inference-cost problem but operates purely in latent embedding space on text sequences. EvoComp is solving a related but distinct challenge: visual tokens carry spatial and semantic information that text tokens don't, so merging them naively destroys grounding. The two papers together suggest the field is converging on a shared pressure point, namely that raw token counts are the primary inference bottleneck, but splitting into separate tracks for text-only versus multimodal pipelines.

The real test is whether EvoComp's accuracy retention holds on established multimodal benchmarks like MMStar or MMMU at compression ratios above 75%. If independent replication shows degradation on fine-grained visual reasoning tasks at high compression, the evolutionary labeling advantage shrinks considerably.